基于图神经网络的轻量化\par 时空交通流预测方法

\justifying Traffic flow prediction technology is an important component of intelligent transportation systems. Currently, spatiotemporal traffic flow prediction based on deep learning has become mainstream. However, deep learning methods often have a large number of parameters. A good prediction method should ensure both accuracy and adapt to the current conditions of edge intelligent hardware, trying to minimize the model's size. However, these two goals are often difficult to balance. Currently, many spatiotemporal traffic flow prediction models ignore the balance between model size and prediction accuracy, which limits their practicality to some extent. In response to this problem, this article proposes a lightweight spatiotemporal traffic flow prediction algorithm: Graph Convolution and Temporal Convolution Network (GCTCN). GCTCN includes a graph convolution module that is improved for traffic prediction, which can effectively model the dependence relationship between long-distance vertices and efficiently extract spatial features of the road network; it also includes a time feature extraction module based on dilated causal convolution, which can extract time features of different scales. By combining the above methods, it can model long- and short-distance spatiotemporal dependence relationships without increasing the number of parameters, effectively improving the prediction ability. Experimental results have shown that this method can balance spatiotemporal efficiency and prediction accuracy.